JEV RECIPES

Take the judgements back from the large model

Classifying, moderating, grading, routing, gating — these produce an enum and a probability, not a paragraph. Below are 5 jobs and 15 ready-made question sets, each runnable on this page with your own text.

LLM-as-a-Judge with JevPutting one large model in charge of grading another is slow, expensive, and the confidence it reports clusters around 95%, so it sorts nothing. Grading is a judgement, not an essay — which is the shape a System One model is built for.3 recipes · runnable here
Content moderation with JevModeration output is a handful of booleans and a disposition, not a paragraph. Run it through a generative model and every item costs you a second of latency, a round of tokens, and some JSON parsing you would rather not maintain.3 recipes · runnable here
Intent classification and ticket triage with JevTriage produces an enum and a priority, not prose. With a generative model you spend the prompt talking it into returning only those values, and then handle the times it improvises anyway. With Jev the answer space is part of the call.3 recipes · runnable here
Tool-call guardrails for agents with JevEvery tool call an agent makes deserves a judgement before it runs: can this go through unattended? Ask a large model and you add a second per step. Write rules and you never finish writing them. What fits here is a probability-bearing boolean in a few hundred milliseconds.3 recipes · runnable here
Semantic routing with JevRouting is where a small spend prevents a large one: pay a fraction of a cent to decide whether this request needs to wake the model that costs a dollar a call. That only works if the decision itself is fast and cheap — otherwise the router becomes the bottleneck.3 recipes · runnable here

Why it is worth switching

We put Jev, GPT-4.1-mini and GPT-5.6 Sol / Terra / Luna through the same 2,390 labelled questions. True/false accuracy 92.5%, level with the newest reasoning models; median latency 456 ms against 1.3–1.7 s; $0.049 against $4.27 for the whole run. The costs: about 5 points behind on multiple choice, and clearly weak at counting.

The full data, a claim-by-claim check of the official numbers, and the limits of the run are in the benchmark report (written in Chinese for now).